Probing for Referential Information in Language Models

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4177–4189 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 4177 Probing for Referential Information in Language Models Ionut-Teodor Sorodoc∗ Kristina Gulordava Gemma Boleda∗ † ∗Universitat Pompeu Fabra †ICREA Barcelona, Spain {firstname.lastname}@upf.edu Abstract Language models keep track of complex lin- guistic information about the preceding con- text – including, e.g., syntactic relations in a sentence. We investigate whether they also capture information beneficial for resolving pronominal anaphora in English. We analyze two state of the art models with LSTM and Transformer architectures, respectively, using probe tasks on a coreference annotated corpus. Our hypothesis is that language models will capture grammatical properties of anaphora (such as agreement between a pronoun and its antecedent), but not semantico-referential information (the fact that pronoun and an- tecedent refer to the same entity). Instead, we find evidence that models capture referential aspects to some extent –though they are still much better at grammar. The Transformer out- performs the LSTM in all analyses, and ex- hibits in particular better semantico-referential abilities. 1 Introduction Neural network-based language models (LMs) have been shown to learn relevant properties of language without being explicitly trained for them. In particular, recent work suggests that they are able to capture syntactic relations to a large ex- tent (Gulordava et al., 2018; Kuncoro et al., 2018; Wilcox et al., 2018). In this paper, we extend this line of research to analyze whether they are able to capture refer- ential aspects of language, focusing on anaphoric relations (pronoun-antecedent relations, as in she- Yeping Wang in Figure 1). Previous work, such as Ji et al. (2017), Yang et al. (2017) and Cheng and Erk (2019), showed that augmenting language models with a compo- nent that uses an objective based on entity or coref- erence information improves their performance at ...he1 was elected to be president of the Peo- ple’s Republic of China, and chairman of the2 Central2 Military2 Commission2. Yeping3 Wang3 was born in Shanghai in 1926. She3 studied in Shanghai Foreign Language Col- lege, and started working in 1949. For a long time, she3... Figure 1: Example from OntoNotes with a window of 60 tokens (as used in our first probe task). Both occurrences of she refer to the same entity as Yeping Wang. Note that not all entity mentions are annotated in OntoNotes –only those that enter into coreference relationships in the document. language modeling. Intuitively, in the example in Figure 1, understanding that the first she refers to Yeping Wang makes words related to studying or working more likely to follow than other kinds of words. That is, referential information helps lan- guage models do their task. The cited work includes explicit coreference guidance; however, since referential information is useful for language modeling, we expect lan- guage models to learn referential information even without explicit supervision. Here we analyze to what extent this is the case. We carry out our analysis using probe tasks, or tasks that check …

Content truncated to 500 words